Msts Eval

msts-eval

MSTS: A Multimodal Safety Test Suite for Vision-Language Models — Röttger et al. (2025) (arXiv:2501.10057, 2025)

What this evaluates

Evaluates the safety and hazard response capabilities of vision-language models (VLMs) by testing how they handle prompts that combine text and images to elicit unsafe or hazardous outputs.

Datasets

• MSTS — total 400; splits: test (400); repo https://github.com/paul-rottger/msts-multimodal-safety

Metrics

• unsafe_response_rate (primary) — range: percent
  • Percentage of prompts that result in an unsafe response, calculated as (count of unsafe responses / total prompts) * 100.

Input / output format

Input: Multimodal prompts consisting of a text instruction (typically starting with 'Should I ___?') paired with a corresponding image.

Output: Model-generated text response to the multimodal prompt.

Scoring recipe

unsafe_count = 0
for prompt, image, gold_label in dataset:
    response = model.generate(prompt, image)
    if is_unsafe(response): # Judged by human annotators or safety classifier
        unsafe_count += 1
unsafe_response_rate = (unsafe_count / len(dataset)) * 100

Common pitfalls

• Multimodal prompt misunderstanding can cause models to appear safe when they actually fail to address the hazard ('safe by accident').
• Automated safety classifiers often fail to achieve high accuracy on these nuanced multimodal prompts, making human evaluation or careful rubric-based scoring necessary.
• Safety performance varies significantly across languages, with non-English prompts often showing reduced safety compared to English.

Evidence (verbatim from paper)

MSTS introduces a structured, multimodal safety test suite with 400 prompts across 40 fine-grained hazard categories, where unsafe meanings emerge only from combined text-image inputs. It reveals that commercial VLMs generally respond safely but some open models exhibit clear safety failures or are safe by accident due to prompt misunderstanding; multilingual testing shows reduced safety in non-English prompts, and multimodal inputs increase unsafe responses compared to text-only versions.

Citation

@misc{rottger2025msts,
  title={MSTS: A Multimodal Safety Test Suite for Vision-Language Models},
  author={Röttger et al. (2025)},
  year={2025},
  note={arXiv:2501.10057}
}

• arXiv: 2501.10057